Hands-On Reinforcement Learning for Games

Summary

In this chapter, we looked beyond DRL and into the realm of AGI, or at least where we hope we are going with AGI. More importantly, though, we looked at what the next phase of DRL is, how we can tackle its current shortcomings, and where it could go next. We looked at meta learning and what it means to learn to learn. Then we covered the excellent learn2learn library and saw how it could be used on a simple deep learning problem and then a more advanced meta-RL problem with MAML. From there, we looked at another new approach to learning using hindsight with HER. From hindsight, we moved to imagination and reasoning and how this could be incorporated into an agent. Then we finished the chapter by looking at I2A—imagination-augmented ...